Autonomy unbridled. Governed by math, not blind faith. https://nitter.cf/t.co/iDkNXZRf5m
Hamilton, Ontario
Joined March 2023
- Tweets1.7K
- Following39
- Followers38.4K
- Likes1.2K
Pinned Tweet
Inference Labs took the stage in Dubai and came away with the audience vote.
Big thanks to the Daos Hub Dubai team for bringing together founders, investors, and operators, and to @colingagich for representing us and presenting the pitch.
A model escaping an evaluation environment is no longer hypothetical. Recent disclosures show exactly why sandboxing alone is not enough.
Containment limits damage. Evidence reconstructs what happened.
Autonomous systems need both.
inferencelabs.com/
AI can now help formally verify security properties across 100+ real software libraries.
That is a big shift: AI can help generate proof, not just code.
At Inference Labs, we think verification should become a normal part of machine-produced work.
inferencelabs.com/
1/ Operational problems are cheapest to solve before they become incidents.
Congestion starts building. Snow coverage leaves a gap. A baggage position slows down.
Computer vision can surface the pattern while there is still time to act.
Computer use is getting cheaper and faster. That means agents can take far more actions for the same budget.
The bottleneck moves again: not “can AI click the button?” but “can we prove which action happened and why?”.
That is the runtime layer Inference Labs is building for.
Low visibility does not have to mean low awareness.
Sertn can keep tracking ferries, cargo vessels, and other marine traffic through haze or poor conditions, helping port teams maintain a clearer operational picture when visibility drops.
1/ A lot of airport events are impossible to recreate. A tow crosses the wrong zone. A stand entry is off-line. Baggage backs up for six minutes.
By the time someone investigates, the physical moment is already gone.
2/ That is why Sertn treats inference as a record, not just an alert.
What happened, when it happened, which model detected it, and the visual evidence behind it can remain available for review later.
sertn.ai/
Air-gapped AI is becoming more relevant in defense, manufacturing, and regulated environments. When data stays on-site, trust cannot depend on a remote service.
Sertn is built for verifiable AI that can operate where the data lives.
sertn.ai/
Self-hosted inference is becoming more common as teams want lower latency, more privacy, and more control. But local deployment should not mean less accountability.
Sertn is built to keep inference verifiable even when it runs close to the edge.
sertn.ai/
1/ Operational dashboards are good at averages.
Airports often care about the exception.
The one aircraft that entered the wrong stand. The one person who stopped moving. The one baggage position that suddenly backed up.
2/ Sertn is designed around those moments.
Models can watch continuously, surface the event when a defined condition breaks, and route it into the next operational step without someone watching every feed.
sertn.ai/
Computer vision is moving from object detection to understanding scenes, relationships, motion, and context. That opens much richer operational workflows.
Sertn is built for that shift from “what is here?” to “what is happening?”.
sertn.ai/
Crowd problems usually become visible only once they are already disruptive.
Sertn can monitor density at choke points and detect unusual stillness or posture, helping airport teams respond to congestion or a person down before someone has to report it manually.
An AI policy can say what a system should do.
Runtime evidence shows what actually happened.
That distinction is becoming important enough that new standards are emerging around independently verifiable AI governance and execution records.
Once computer vision can trigger actions, the useful question changes. It is no longer only “what did the model detect?” but “what happened because of it?”.
Sertn helps connect inference, action, and evidence in one workflow.
1/ Traditional enterprise auditing is retrospective.
Collect logs, interview people, reconstruct events, then decide whether controls behaved as expected.
AI operating at machine speed makes that model increasingly expensive.
Most enterprise integrations expose an API. The next generation may also expose evidence.
Input accepted. Model executed. Output produced. Proof attached.
At Inference Labs, we think verification can become an interface between AI vendors, customers, auditors, and machines.
1/ Enterprise AI often has two things it cannot freely expose:
customer data and proprietary models.
That creates an awkward trust problem. A third party may need assurance that the right computation happened without being allowed to inspect everything inside it.